Karma browser disconnect leaves pending specs unassigned

Handling Browser Disconnects in Karma

Karma does not automatically reassign pending specs from a disconnected browser to other remaining instances. Once a spec file is sharded and assigned to a specific browser instance via the coordination layer, that assignment is static for the duration of that shard's lifecycle.

Execution Behavior and Completion

When a browser disconnects mid-run, the following behaviors typically occur:

• Spec Status: Specs assigned to the disconnected browser that have not yet reported a result generally remain in an unexecuted state or are eventually marked as failed. They are not redistributed to healthy browsers.
• Run Completion: A run is considered complete when all assigned shards have either reported a result or the connection is officially timed out. The runner does not trigger a re-balance of the remaining workload.
• Concurrency: While config.concurrency limits the number of simultaneous instances, it does not provide a dynamic failover mechanism. If an instance crashes, the concurrency slot may open, but the lost work is not rescheduled.

Likely Technical Explanation

This behavior stems from Karma's orchestration model. The socket.io layer manages the communication between the server and the browser, but the distribution of spec files is handled as a partitioning task rather than a dynamic queue. Because framework adapters handle retries at the spec level inside the browser, they cannot recover from a total process failure or socket disconnection.

Verification and Mitigation

To verify this behavior in your specific environment, you can use a scoped test:

1. Configure a suite with multiple shards and concurrency > 1.
2. Manually terminate a browser process (e.g., kill -9 on a headless Chrome PID) during execution.
3. Observe the console output to confirm that the specs assigned to that PID are not picked up by other active browsers.

To manage the impact of disconnects, review your browserDisconnectTimeout setting in the Karma configuration to ensure the runner fails fast rather than hanging indefinitely.

Diagnostic Detail Required

Are you using a custom browser launcher or a containerized environment (like Docker/Kubernetes) to spawn these instances? This may affect how quickly the socket.io layer detects the disconnect versus a standard local process crash.